Abstract
<title>Abstract</title> <p>Neural architecture search (NAS) is conventionally treated as a data-centre workload. The GaaR framework (Goal-Aware Adaptive Regulation) was introduced to contest that assumption, demonstrating evolutionary architecture search for convolutional networks entirely on a consumer AMD Radeon RX 5700 (8 GB), reaching 94.36% CIFAR-10 accuracy in 45 minutes with full experiment provenance in a locally queryable SQLite log. This paper reports that the same engine transfers, without structural modification, to a second modality: autoregressive transformer language models. We couple GaaR's exploit-best/explore-mutation loop to an unmodified from-scratch PyTorch transformer implementation and run the search through DirectML on the same consumer GPU, a device with no CUDA and no official ROCm support. In an illustrative eight-generation search over embedding width, head count, depth, context length, batch size, and learning rate (fixed budget of 300 optimization steps per candidate), the engine improved character-level validation loss on the Tiny Shakespeare corpus from 2.171 (seed) to 2.021 (perplexity 7.5) with a 0.8 M-parameter champion, completing the entire search in 3 minutes 16 seconds of wall time. A ten-seed controlled comparison against uniform random search finds statistical parity at this eight-evaluation budget (mean best validation loss 2.047 vs 2.042; five wins each head-to-head) — while the identical greedy rule dominates random search on a 6,000-configuration physics-simulation reference landscape (exact optimum located in 70% vs 0% of repeats). Together these results delimit where regulated exploitation pays: not within trivially small budgets on smooth landscapes, but on larger structured ones. The search also exhibits an instructive budget effect: under a fixed step budget, selection favours wide-shallow, short-context genomes whose higher step throughput converts the budget into more effective learning. We describe the DirectML portability envelope (operator fallbacks, no mixed precision), release all code and the generation logs, and argue that cross-modal NAS on commodity hardware is a practical foundation for reproducible, energy-modest architecture research.</p>